Detecting Mislabeled Annotations with Superannotate

Several ML techniques explore the problem of detecting mislabeled annotations in noisy data. SuperAnnotate allows users to detect risky annotations and import them to the platform via the instance-based Approve/Disapprove functionality.

As a result, risky annotations are automatically suggested to be disapproved and a significant portion of images with clean annotations exit the QA process.

Installation
Running the code
Tutorials
License

Installation

Our Colab Notebook has step by step instructions on installation of mislabel detection package and its dependencies.

Requirements

Linux or macOS with Python ≥ 3.6
PyTorch ≥ 1.4 and torchvision that matches the PyTorch installation.
Detectron2 and its dependencies

Run this commands to install Superannotate Python SDK and mislabel detection package

pip install superannotate
pip install git+https://github.com/superannotateai/qa-automation

Running the code

We implemented a simple method to detect mislabeled bounding box annotations in data using Fast-RCNN softmax logits.

The code block below detects mislabeled and missing bounding box annotations in given noisy dataset.

import autoqa

configs = autoqa.train_fastrcnn_on_noisy_dataset(dataset_root, batch_size = 4, num_iter = 200)

autoqa.detect_mislabeled_annotations(configs)
autoqa.detect_missing_annotations(configs)

It generates the annotation output in SuperAnnotate’s format with the error field enabled for noisy annotations.

[{
  "type": "bbox",
  "classId": 9,
  "className": "cat", 
  "points": {"x1": 130.0, "x2": 251.0, "y1": 227.0, "y2": 315.0},
  "error": true
  }]

Detected risky annotations can be imported to the SuperAnnotate platform via Superannotate Python SDK.

import superannotate as sa

sa.init(authorization_token_path)
sa_project = sa.search_projects(project_name)[0]

sa.upload_annotations_from_folder_to_project(sa_project, mislabeled_annotations_root)
sa.upload_annotations_from_folder_to_project(sa_project, missing_annotations_root)

Tutorials

See our Colab Notebook as a step-by-step guide for detecting mislabeled annotations in PASCAL VOC dataset.

License

This project is licensed under the Apache License - see the LICENSE.md file for details.

AnnotationSoftware / qa-automation